首页|DMFI-Net : Dual-branch multi-scale feature interaction network integrating transformer and CNN-Wavelet for image classification of colorectal polyps
DMFI-Net : Dual-branch multi-scale feature interaction network integrating transformer and CNN-Wavelet for image classification of colorectal polyps
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NETL
NSTL
Elsevier
Risk classification of colorectal polyps in endoscopic images is crucial for improving clinical diagnostic accuracy and reducing colorectal cancer mortality. However, this task faces significant challenges due to blurred boundaries between polyps and the colorectal wall, varying intra-class scales, and high inter-class similarity. To address these issues, we propose a Dual-Branch Multi-Scale Feature Interaction Network (DMFINet). In this network, the Transformer branch captures global features, enhancing the ability to model long-range dependencies, particularly by focusing on subtle polyps through a Multi-Head Fusion Self-Attention Module (MFSM). Simultaneously, the CNN-Wavelet branch extracts local detail features, using Discrete Wavelet Transform (DWT) instead of traditional pooling and convolution operations to better preserve edge and texture information. Additionally, we designed an Interactive Connection Module (ICM) to effectively aggregate features from both branches in an interactive manner, thereby reducing the impact of high inter-class similarity and intra-class variability. We evaluated DMFI-Net on a private colorectal polyp dataset and the public Kvasir dataset, achieving classification accuracies of 75.21% and 87.67%, respectively, which is higher than the compared state-of-the-art colorectal polyp classification algorithms.
NETL
NSTL
Elsevier
